Assessing data variability in groundwater quality monitoring of contaminated sites through factor analysis and multiple linear regression models

Monitoring of long-term contaminant concentrations trends is essential to verify that attenuation processes are effectively occurring at a site. However, monitoring data are often affected by extreme variability which prevents the identification of clear concentration trends. The variability is higher in long-screened monitoring wells, which are currently used at many contaminated sites, although it has been known since the 1980s that monitoring data from long-screened wells can be biased. Understanding the factors that may influence the variability of monitoring data is pivotal. To this end, following hydrochemical conceptual modelling using a multi-method approach, the variability of hydrocarbon concentrations from fully screened monitoring wells was assessed over eleven years at a former oil refinery located in Northern Italy. The proposed methodology combined factor analysis with multiple linear regression models. Results pointed out a higher variability in hydrocarbon concentrations at the plume fringe and a lower variability at the plume source and core. 44-46 % of the total variability in measured hydrocarbon concentrations is due to "intrinsic plume heterogeneity", related to the three-dimensional structure of a contaminant plume, which becomes thinner at the edge, creating a vertical heterogeneity of redox conditions at the plume fringe. This variability, expressed as increasing concentrations of sulfate and decreasing concentrations of methane, represents a background variability that cannot be reduced by improving sampling procedures. The remaining 56-54 % of the total variability may be due to the non-standardization of some purging and sampling operations, such as pump intake position, purging and sampling time/flow rates and variations in the analytical methods. This finding suggests that monitoring improvements in fully screened wells by standardizing all purging/sampling operations or using sampling techniques that can reduce the actual screen length (e.g., packers or separation/dual pumping techniques) would reduce data variability by more than half.